Introduction to Deep Learning
Angelica Sun
(adapted from Atharva Parulekar, Jingbo Yang)
Overview
● Motivation for deep learning
● Convolutional neural networks
● Recurrent neural networks
● Transformers
● Deep learning tools
But we learned multi-layer perceptron in class?
Expensive to learn. Will not generalize well.
Does not exploit the order and local relations in the data!
64x64x3=12288 parameters We also want many layers
Convolutional NN
Image
What are areas of deep learning?
Recurrent NN
Sequential Inputs
Deep RL
Graph NN
Networks/Relational
Transformers
Parallelized Sequential Inputs
Control System
Starting from CNN
Convolutional Neural Network
Let us look at images in detail
Filters in traditional Computer Vision
Image credit: https://home.ttic.edu/~rurtasun/courses/CV/lecture02.pdf
Learning filters in CNN
Why not extract features using filters?
Better, why not let the data dictate what filters to use?
Learnable filters!!
Convolution on multiple channels
Images are generally RGB !!
How would a filter work on a image with RGB channels?
The filter should also have 3 channels.
Now the output has a channel for every filter we have used.
Parameter Sharing
Lesser the parameters less computationally intensive the training. This is a win win as we are reusing parameters.
Translational invariance
Since we are training filters to detect cats and the moving these filters over the data, a differently positioned cat will also get detected by the same set of filters.
Visualizing learned filters
Images that maximize filter outputs at certain layers. We observe that the images get more complex as filters are situated deeper
How deeper layers can learn deeper embeddings. How an eye is made up of multiple curves and a face is made up of two eyes.
A typical CNN structure:
Image credit: LeCun et al. (1998)
Convolution really is just a linear operation
In fact convolution is a giant matrix multiplication.
We can expand the 2 dimensional image into a vector and the conv operation into a matrix.
SOTA Example – Detectron2
How do we learn?
Instead of
They are “optimizers”
● Momentum: Gradient + Momentum
● Nestrov: Momentum + Gradients
● Adagrad: Normalize with sum of sq
● RMSprop: Normalize with moving
avg of sum of squares
● ADAM: RMsprop + momentum
Mini-batch Gradient Descent
Expensive to compute gradient for large dataset Memory size
Compute time
Mini-batch: takes a sample of training data
How to we sample intelligently?
Is deeper better?
Deeper networks seem to be more powerful but harder to train.
● Loss of information during forward propagation
● Loss of gradient info during back propagation
There are many ways to “keep the gradient going”
One Solution: skip connection
Connect the layers, create a gradient highway or information highway.
ResNet (2015)
Image credit: He et al. (2015)
Initialization
Can we initialize all neurons to zero?
If all the weights are same we will not be able to break symmetry of the network and all filters will end up learning the same thing.
Large numbers, might knock relu units out.
Relu units once knocked out and their output is zero, their gradient flow also becomes zero.
We need small random numbers at initialization.
Variance : 1/sqrt(n) Mean: 0
Popular initialization setups
(Xavier, Kaiming) (Uniform, Normal)
Dropout
What does cutting off some network connections do?
Trains multiple smaller networks in an ensemble.
Can drop entire layer too!
Acts like a really good regularizer
More tricks for training
Data augmentation if your data set is smaller. This helps the network generalize more.
Early stopping if training loss goes above validation loss.
Random hyperparameter search or grid search?
CNN sounds like fun!
What are some other areas of deep learning?
Recurrent NN Sequential data
Convolutional NN
Deep RL
Graph NN
We can also have 1D architectures (remember this)
CNN works on any data where there is a local pattern
We use 1D convolutions on DNA sequences, text sequences, and music notes
But what if time series has causal dependency or any kind of sequential dependency?
To address sequential dependency?
Use recurrent neural network (RNN)
Step output
Latent Output RNN Cell
Input at one time step
Unrolling an RNN
The RNN Cell (Composed of Wxh and Whh in this example) is really the same cell. NOT many different cells like the filters of CNN.
How does RNN produce result?
Evolving “embedding”
I love CS !
Result after reading full sentence
2 Typical RNN Cells
Store in “long term memory” Response to current input Reset gate Update gate
Long Short Term Memory (LSTM) Gated Recurrent Unit (GRU)
Response to current input
Recurrent AND deep? Taking last value
Pay “attention” to everything
Stacking
Attention Model
Transformer – Attention is All You Need!
Originally proposed for translation.
Encoder computes hidden representations for each word in the input sentence
Applies self attention.
Decoder makes sequential prediction similar as in RNN
At each time step, it predicts the next word based on its previous predictions (partial sentence). Applies self attention and attention on encoder outputs.
Transformer – Attention is All You Need!
The dot product in softmax below computes how each word of sequence 1 (Q) is influenced by all the other words in the sequence 2 (K).
Considering the different importance, we computed a weighted sum of the information in the sequence 2 (V) to use in computing the hidden representation of sequence 1.
Transformer – Attention is All You Need!
The dot product in softmax below computes how each word of sequence 1 (Q) is influenced by all the other words in the sequence 2 (K).
Considering the different importance, we computed a weighted sum of the information in the sequence 2 (V) to use in computing the hidden representation of sequence 1.
Transformer – Attention is All You Need!
Multiple heads!
— Similar as how you have multiple filters in CNN
Loss of sequential order?
— Positional encoding! (often use sine waves)
Examples of attention scores from two different self-attention heads.
References:
https://arxiv.org/pdf/1706.03762.pdf https://medium.com/inside-machine- learning/what-is-a-transformer- d07dd1fbec04 https://towardsdatascience.com/transfor mers-141e32e69591 https://towardsdatascience.com/transfor mers-explained-visually-part-2-how-it- works-step-by-step-b49fa4a64f34
SOTA Example – GPT3
SOTA Example – GPT3
SOTA Example – DALLE
More? Take CS230, CS236, CS231N, CS224N
Convolutional NN Image
Recurrent NN Time Series
Graph NN Networks/Relational
Deep RL Control System
Not today, but take CS234 and CS224W
Convolutional NN Image
Recurrent NN Time Series
Graph NN Networks/Relational
Deep RL Control System
Tools for deep learning
Specialized Groups
Popular Tools
$50 not enough! Where can I get free stuff?
Google Colab
Free (limited-ish) GPU access
Works nicely with Tensorflow
Links to Google Drive Register a new Google Cloud account
=> Instant $300??
=> AWS free tier (limited compute) => Azure education account, $200?
Azure Notebook Kaggle kernel??? Amazon SageMaker?
To SAVE money
CLOSE your GPU instance
~$1an hour
Good luck!
Well, have fun too 😀